From Convolution to Attention: Transformer-Based Modeling for Multi-Day Wildfire Spread Forecasting

Wildfire prediction and management is a critical component of environmental risk assessment, especially in the face of increasing climate volatility. Recent advances in remote sensing and artificial intelligence have enabled more precise forecasting of wildfire spread dynamics. However, existing deep learning models often struggle to capture complex spatiotemporal interactions and translate predictions into actionable decisions. This paper addresses this gap by integrating transformer-based prediction with a BDI (Belief–Desire–Intention) reasoning framework for real-time wildfire response. The study leverages four open-access datasets: WFED, FIRMS, Sentinel Hub, and a custom ERA5 + MODIS + SRTM fusion, encompassing fire masks, vegetation indices, meteorological parameters, and topographic features. The proposed methodology constructs spatiotemporal tensors with learnable positional encodings, processed through a multi-head attention transformer to predict burnt area masks and spread direction vectors. These predictions inform a role-specific BDI reasoning layer to generate adaptive action plans. The novelty lies in combining attention-driven deep learning with symbolic goal-based decision modelling in an operational wildfire management context. We evaluated the model performance using F1-score, IoU, MAE, directional accuracy, and AIC-based feature importance. Results show the proposed Transformer + BDI system significantly outperforms prior CNN-based approaches, achieving an F1 score of 0.75 and reducing prediction MAE to 7.4 km2 across datasets, while offering interpretable and actionable wildfire response insights.